Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

I'm not sure why anyone would even do that in the first place, LLama doesn't generate synthetic data that would be even remotely good enough. Even GPT 3.5 and 4 are already very borderline for it, with lots of wrong and censored answers. And at best you make a model that's as good as LLama is, i.e. not very.


Instruction-tuning is the obvious use case. That much has nothing to do with subjectivity, alignment or censorship, it's will-you-actually-show-this-as-JSON-if-asked.


That's tuning llama which is allowed from what I understand. Otherwise why release it at all, it's not very functional in its initial state anyway. What that applies to is using llama outputs to train a completely new base model which makes no practical sense.

As for generating jsons, that's more of a inference runtime thing, since you need to pick the top tokens that result a valid json instead of just hoping it returns something that can be parsed. On top of extensive tuning of course.




Consider applying for YC's Fall 2026 batch! Applications are open till July 27.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: